As I was leaving Tom Mitchell’s office, he says to me in the kind of hurried speech of a brilliant individual who has perhaps made the statement before:
“Computers are already better than us at playing chess, but we are still better at recognizing a photo of our parents or children.”
Many consider Tom Mitchell, chair of the machine learning department at Carnegie Mellon University, to be a leading pioneer and expert in the field; and I just met with him to get a better understanding of machine learning, its limits, and its future.
Some major takeaways include:
- We should not be afraid of machine learning replacing humans, at least in the near term
- Even the area of “unsupervised learning” still requires humans to tell the computer what relationships to tease out of data
- The lack of understanding possessed by businesses and marketers about machine learning consistently causes them to come to machine learning experts, such as Dr. Mitchell, asking these computer scientists to make sense of these large datasets. These are under specified and arguably therefore useless questions for which machine learning is of little use. What is always needed is context, a hypothesis you wish to test, and an ability to gather/capture useful datapoints.
At the end of the day, humans are still needed to perform one very critical function: defining the dependent variable and the range of possible independent variables that explain/affect that dependent variable.